In 1817, in his document “An Essay on the Shaking Palsy”, James Parkinson was the first to describe Parkinson’s disease (PD) as a neurological syndrome, characterized by a shaking palsy [1,2]. Later in 1872, the French neurologist Jean-Martin Charcot, described this disease more precisely and distinguished bradykinesia from other tremendous disorders by examining a large number of patients and suggested the use of the term “Parkinson’s disease” for the first time [3,4]. Further studies were made until Brissaud and Meige [5] identified damage of the substantia nigra as the main cause of PD. This damage leads to a range of symptoms including rigidity, balance impairment, rest tremor, and slowness of movement [6].

In addition to these motor symptoms, the voice is also affected. Voice and speech impairment is a typical symptom of PD that occurs in most patients [7]. Gradual deterioration of communication skills in patients with PD is considered to be a significant cause of disability [7]. Ho et al. [8] found that 147 PD patients out of 200 had speech impairment, and participants showed a gradual deterioration of speech characteristics. Traditional diagnosis of PD is costly, and it can take from hours to a few days to be performed. Consequently, evaluating the consistency of the voice and recognizing the triggers of its deterioration in the sense of PD based on phonological and acoustic signals is essential to improving PD diagnosis. Furthermore, developing a smart system based on machine-learning (ML) techniques able to detect this disease in an early stage will reduce the number of clinical visits for examinations and the workload of clinicians [9].

For example, Little et al. [10] presented a system that detects dysphonia by discriminating between healthy controls (HC) and PD participants using a dataset of 195 records collected from 31 patients, of which 23 were diagnosed with PD. They extracted both time domain and frequency domain features from the records and achieved a classification accuracy of 91.4% using 10 highly uncorrelated measures and the support vector machine (SVM) technique. Benba et al. [11] used a dataset consisting of voice samples of 17 PD patients and 17 HCs recorded using a computer’s microphone. They extracted 20 Mel-frequency cepstral coefficients (MFCC) and achieved a classification accuracy of 91.17% using linear SVM with 12 coefficients. Hemmerling et al. [12] used an original dataset consisting of 198 records of 33 PD patients and 33 HCs. They extracted several acoustic features, and applied principal component analysis (PCA) for feature selection and used a linear SVM classifier, which achieved an accuracy of 93.43%.

In this paper, we describe our methodology that analyzes raw audio recordings collected using smartphones to create accurate predictive models. As previously mentions, several studies have been conducted on this subject, but our methodology differs in many ways. First, we used a large dataset, which consisted of 18,210 recordings, where 9,105 were obtained from 453 patients with PD and 9,105 were obtained from 1,037 HCs. To the best of our knowledge, this is the largest cohort data used in a clinical application. Second, instead of extracting only time, frequency, or cepstral domain features from the recordings, we used a combination of the three domains to create highly accurate predictive models. Our final dataset consisted of 80,594 instances and 138 features as well as a class variable. We applied two feature selection methods, analysis of variance (ANOVA) and least absolute shrinkage and selection operator (LASSO), to select the best subset of features. Then we compared our method with various state-of-the-art and newer ML techniques, namely, linear SVM, K-nearest neighbor (KNN), random forest (RF), and extreme gradient boosting (XGBoost). A maximum accuracy of 95.78% was achieved using XGBoost on unseen data.

The remainder of the paper is organized as follows. We describe our method in detail in Section II, Section III presents the results, and Section IV discusses the findings.

This paper presented our method which we used to classify PD patients and distinguish them from HCs using 18,210 smartphone recordings by creating a dataset of 80,594 samples with 138 features and a class variable. The LASSO feature selection method outperformed ANOVA with almost the same number of features. From Supplementary Table S3, we conclude that age and gender features are highly ranked using both methods; it is known that PD is seen in people aged over 50, and it affects males more than females. Furthermore, energy entropy, spectral spread, and MFCC coefficients are highly ranked using both methods, which indicates the importance of extracting time, frequency, and cepstral domain features, in addition to the age and the gender of participants in classifying this disease.

From the same table, we can notice that there is a difference in the ranking of features between ANOVA and LASSO because each technique implements a different approach. ANOVA analyses the relationship between each feature and the class variable separately and assigns a test score to each feature. Then all the test scores are compared, and the features with top scores are selected (K = 10, 20, 30). On the other hand, LASSO regularization adds a penalty to the different parameters of the model to avoid overfitting. This penalty is applied over the coefficients that multiply each of the features. Thus, the L1 technique analyses all the features at once. In addition, LASSO has an important property of shrinking down to zero unimportant features, which depends on the chosen C parameter. For this reason, in Supplementary Table S3, there are different features for each C value in LASSO and the same ranked features using ANOVA with regards to the chosen K value. With this combination of features, XGBoost outperformed the remaining classifiers with an accuracy of 95.31% using 80% of the data (Table 5) while predicting new cases on unseen data with an accuracy of 95.78% (Table 6).

Several studies [10–12] have reported high classification accuracy using SVM in the range of 91% to 93%, as seen in Table 7. However, in Tables 3–6, we note that SVM was the least accurate, with a maximum accuracy of 76.47% using 138 features. This is attributed to the fact that these studies used a small number of recordings, and small datasets. Furthermore, we found that a regular SVM takes time to fit the data (approximately 40 minutes). For this reason, we used linear SVM, which is optimized for large datasets, which limited our chances to test various other kernels (RBF, poly, and sigmoid).

Singh and Xu [19] used the same dataset that we used and achieved an accuracy of 99% using MFCC coefficients, L1-based feature selection, and an SVM classifier with an RBF kernel using 1,000 samples. The problem is that those 1,000 recordings were chosen randomly from an unbalanced database of 65,022 recordings, where 14% of participants were diagnosed with PD and 86% were healthy controls (claimed in their paper). Randomly choosing 1,000 recordings from an unbalanced set of recordings may have introduced an unbalanced set of 1,000 recordings. Moreover, those recordings were chosen without taking into account the medication time point; therefore, their dataset may have included some recordings of patients after they had taken PD medication. We avoided this in our cohort selection phase, as seen Figure 1, compared to our dataset where we made a 50/50% split of the recordings, where 9,105 were obtained from PD patients, and 9,105 were obtained from HCs. Furthermore, relying only on accuracy as a metric to assess the performance of the classifiers is not sufficient in medical diagnostic. Adding other metrics, such as sensitivity (which measures the proportion of PD patients that were correctly classified as having PD), and specificity (which measures the proportion of HCs that were correctly classified as not having PD) will give a better estimate of the performance of the classifiers. In our case, our highest classifier achieved an accuracy of 95.78%, a sensitivity of 95.32%, and a specificity of 96.23% with an F1-score of 95.74%. In the case of an unbalanced dataset (which is not clear in their case), relying only on accuracy may results in a high accuracy if one class is outnumbered; thus, introducing other metrics is important. Hence, we believe that our approach is more accurate and precise for classifying PD even if Singh et al. [19] achieved a higher accuracy which could be affected by their methodology choices.

In conclusion, we proposed a method to classify PD using a large sample of smartphone recordings as a sustained phonation of /a/ for 10 seconds. These recordings were then processed to extract multiple domain features in addition to demographic parameters to create an original dataset that was subjected to various ML techniques after data cleaning, normalization, and feature selection. We have demonstrated the importance of using these features to precisely classify PD with an accuracy of 95.78% using XGBoost. The main objective of this work was to build a smart framework based on ML techniques capable of distinguishing between PD patients and HCs using voice as a disease biomarker. As a future work, we aim to develop an mHealth system capable of implementing these ML techniques to speed up diagnosis time and to integrate it with conventional clinical methods.